Two applications for the case of Tajikistan and a simulation study
Kristina Meier
A thesis submitted to the Georg August University of Goettingen in fulfilment of the requirements for the degree of
Doctor of Philosophy
February 5, 2013
First of all I would like to thank Prof. Klasen for his support over the years.
Furthermore, I thank Prof. Lay and Prof. Kneib for helpful comments on this dis- sertation. Finally, my gratitude goes to everyone who made the time in Goettingen pleasant and memorable.
Introduction vii 1 Low-skilled labor migration in Tajikistan:
Determinants and effects on expenditure patterns 1
1.1 Introduction . . . 1
1.2 The case of Tajikistan . . . 3
1.3 Data . . . 4
1.4 Modeling determinants of labor migration . . . 5
1.5 Econometric methodology . . . 7
1.6 Descriptives . . . 8
1.7 Results . . . 10
1.8 Conclusion . . . 23
2 Employment and the financial crisis in Tajikistan 25 2.1 Introduction . . . 25
2.2 Background on Tajikistan . . . 27
2.3 The impact of crises on labor market outcomes: Some theory and empirical evidence . . . 28
2.4 Data . . . 29
2.5 Methodology . . . 30
2.6 Results . . . 36
2.7 Conclusion . . . 50 ii
3 A sensitivity check of the recursive bivariate probit model under various misspecifications with and without a valid instrument 53
3.1 Introduction . . . 53
3.2 The recursive bivariate probit model . . . 55
3.3 Simulation setup . . . 58
3.4 Results . . . 63
3.5 Conclusion . . . 65
Appendix 69
Bibliography 88
2.1 Transition probabilities by age . . . 39
2.2 Transition probabilities by expenditure quintile . . . 39
3.1 Employment in the manufacturing sector 2007-2009 . . . 70
3.2 Employment in the agricultural sector 2007-2009 . . . 70
3.3 Employment in the non-production sector 2007-2009 . . . 70
3.4 Bivariate normal distribution . . . 71
3.5 Bivariate t distribution . . . 71
3.6 Bivariate lognormal distribution . . . 72
3.7 Bivariate lognormal-t distribution . . . 72
iv
1.1 Proportion of HHs with and without migrants . . . 8
1.2 Gender distribution among migrants . . . 8
1.3 Proportion of secondary education or higher among migrants . . . . 9
1.4 Activity prior to migration . . . 9
1.5 Country of destination . . . 9
1.6 Marginal effects of the probit model . . . 12
1.7 Average expenditure shares for HHs with and without migrants . . 13
1.8 Results of 2SLS regression on household expenditure shares . . . 15
1.9 Mean comparison of remittances (in Tajik Somoni) between recent and long-term migrants . . . 18
1.10 Time effect of migration on household expenditures . . . 20
1.11 Results of 2SLS regression on household pc income . . . 22
2.1 Employment in the different categories in 2007 and 2009 . . . 37
2.2 Transition probabilities for domestic labor outcomes . . . 38
2.3 Multinomial Probit Model: Labor outcomes in 2009 if inactive in 2007 . . . 40
2.4 Multinomial Probit Model: Labor outcomes in 2009 if unemployed in 2007 . . . 41
2.5 Multinomial Probit Model: Labor outcomes in 2009 if wage-employed in 2007 . . . 42
2.6 Multinomial Probit Model: Labor outcomes in 2009 if self-employed in 2007 . . . 43
2.7 Number of migrants per household in 2007 and 2009 . . . 46 v
2.8 Possible indicators for restriction of labor migration in the host country . . . 47 2.9 Probit model: Determinants of labor migration in 2009 . . . 48
The estimation of impact, i.e. causally attributing outcomes to some influence, is a central problem in development economics. The question of measuring how effective (for example) a project intervention was is of paramount interest for re- searchers, donors and policy makers alike. The results determine future program design, as well as further financial support. Gauging the true impact of an in- tervention is far from trivial, and might depend crucially on the method used.
In general, a move from qualitative to quantitative evaluation techniques can be observed, meaning that is has become more or less standard to base the measure- ment of impact on extensive data, rather than peer group interviews and anecdotal evidence. With the increasing availability of tailor made, as well as multi-purpose survey data, the quantitative approach becomes more and more feasible. The choice of the appropriate method for analyzing the impact of treatment, however, remains crucial and is the subject of extensive debate and research.
The term ”treatment” can be quite broadly defined here, meaning for example some project intervention, such as vaccination programs, building of new roads or providing access to clean water. In addition to that, treatment could also be a household choice, such as sending a labor migrant abroad, or, even more generally, some macroeconomic shock, for example a raise in tariffs, or the occurence of a financial crisis. In all of these cases, the econometric problem at the core is the same: Unless participation in the treatment, however defined, was truly random, comparison between the outcomes of the treated and non-treated most likely suffers from bias due to systematic differences between the two groups. The (impossible) solution to the problem would be to observe the treated at the same point in time, once with and once without having received treatment. Formally, the problem can be expressed like this, as the formula for the average treatment effect on the
vii
treated (ATT):
AT T =E[Y1|T = 1]−E[Y0|T = 1]
whereY1 denotes the outcome with treatment, andY0 is the respective outcome without treatment. T is a binary variable indicating the receipt of treatment. It becomes immediately clear that the second term, E[Y0|T = 1] is unobservable, since no outcomes without treatment for those actually treated can be obtained.
This hypothetical scenario is called the counterfactual. Thus, the problem of impact estimation can be characterized as a missing data problem.
An ever-increasing analytical toolbox to tackle this problem is available. Gener- alizing, it can be said that one tries to proxy the counterfactual as best as possible by instead using E[Y0|T = 0], the outcome of non-participants (called control group). The difference between this proxy and the hypothetical, true counterfac- tual is referred to as (selection) bias:
Bias=E[Y0|T = 1]−E[Y0|T = 0]
Minimizing this bias is at the core of each impact evaluation method.
Two main groups of impact evaluation methods exists, full randomization, as well as quasi-experimental designs. If assignment to treatment can be said to have been truly random, the problem of bias as outlined above averages out, and a mere comparison of mean outcomes of treatment and control group gives the de- sired impact estimate. While this is arguably the most credible method to avoid bias, fully randomized set ups are not easy to come by, since more often than not possible future impact evaluations are not taken into account when planning a project intervention, meaning that the researcher has to make do with already ex- isting, non-randomized data. There are also cases, where randomization of project participation is not feasible because of ethical concerns. How is one, for example, to motivate the random exclusion of children from a vaccination program (pro- vided that sufficient funds exists to vaccinate the entire population in question)?
Also, randomized participation might be impossible for technical reasons, as is the case with most infrastructure projects, such as building waterpipes or new roads. Finally, household choices, which are not part of an administered program,
such as labor migration, obviously cannot be randomized. In situations like these, quasi-experimental methods are therefore the only way to try to estimate impact.
Roughly, these methods can be classified according to four main groups, which are briefly outlined in the following section.
0.0.1 Different quasi-experimental methods
The first group tries to remedy bias by matching on observables. This means that the missing data problem is approached by constructing a control group which resembles the treated as closely as possible with regard to observable characteris- tics. Probably the most prominent example of this technique is propensity score matching, as introduced by Rosenbaum and Rubin, 1983. This approach has the great advantage of reducing the matching dimension to the so-called propensity score, an estimate of the probability to receive treatment on the basis of the chosen observable characteristics. An obvious drawback of matching on observables is the fact that the treatment decision might be influenced by hidden factors, or factors for which data could theoretically be collected, but so far do not exist.
The second group of impact evaluation methods can handle this problem when two basic conditions are fulfilled: Firstly, the researcher has to have data for the treated and the untreated both before and after the treatment was administered.
Secondly, the bias-causing unobserved differences (or heterogeneity) between the two groups must be constant over time. In this case, treatment impact can be es- timated by simply differencing before and after outcomes across both the treated and the untreated (also called double differencing, or DID for short).1
The third group of methods tries to make use of existing design features of project implementation. An example for this is the so-called regression discontinuity de- sign approach (RDD). This is a valid tool if treatment depends on some arbitrary threshold, e.g. only farmers who own less than a certain amount of land are eligible for a micro credit. The idea behind RDD now is to compare outcomes between those individuals just above and just below this cut-off, assuming that they are essentially the same in terms of all other relevant characteristics, such as wealth,
1One prominent example is the study of Pitt and Khandker, 1998, using DID to estimate the effect of Grameen bank micro credits and gender on various indicators of living standards in Bangladesh.
socio-economic status etc.2 While this approach is quite convincing, it can of course only be applied if such a threshold for treatment assignment exists, which might often not be the case.
The forth group of methods relies on some sort of exclusion restriction or instru- ment to estimate impact.3 Here a selection equation determines the treatment participation, while a second estimates the outcome. Such models are either esti- mated using a two-step approach, or simultaneously via maximum likelihood. To obtain identification, the selection equation usually contains at least one variable which strongly influences treatment, but has no direct effect on outcome (and is therefore missing in the outcome equation). This variable(s) is called the exclu- sion restriction or instrument. Such methods have the advantage that they do not require the bias-causing heterogeneity to be constant over time, or to be captured solely by observable factors. However, they suffer from another, substantial weak point, namely the exogeneity assumption of the instrument. As already indicated above, in order to be valid an instrument needs to be both relevant (i.e. strongly determining the treatment decision), as well as exogenous, meaning that it must be uncorrelated with the error term of the outcome equation. While the first condition is easily verified, for example with the F-test of the first-stage regression, no true test exists for exogeneity. In some situations, one can indeed find plausible instru- ments. An example for this is the so-called ”intention-to-treat” approach, where the possibility to receive treatment was randomized (e.g. by giving out vouchers), but the actual treatment decision was made by the respective individual. In most cases, however, exclusion restrictions are less convincing, and ”proof” for their validity relies on the persuasive powers of the researcher. Needless to say, this is an unsatisfactory situation, especially when considering how widely used methods based on exclusion restrictions are in impact evaluation studies. A solution to this dilemma presents the so-called ”identification by functional form”-assumption. As the name suggests, identification is obtained by relying on functional form, rather than by including an instrument in the first stage. The obvious drawback of this approach are possible model misspecifications, since the true functional form is
2See for example Angrist and Lavy, 1999, who use RDD to estimate the effect of class size on scholastic achievement in Israel.
3See Angrist and Krueger, 2009 for an overview of the history and applications of the IV technique.
generally unknown.
The following three chapters all deal with some aspect of the evaluation prob- lem. The first chapter is an impact evaluation, methodologically speaking, how- ever, not of an intervention or a program. Here, the impact of the household decision, namely sending a labor migrant abroad, is estimated using a standard IV approach. The second chapter is not an impact evaluation in the true sense of the word. It aims to quantify the effect of the recent financial crisis, i.e. a macroeconomic shock. This means that there is no control group, since the crisis hit worldwide. This problem is approached using a heterogeneous exposure vari- able.
The last chapter is more theoretical in nature. It tries to answer the question of whether to use an (uncertain) instrument, or rather rely on the functional form assumption when using a recursive bivariate probit model to estimate the impact of a binary treatment on a binary outcome. Model performance in the face of various misspecifications of the error term is simulated with and without a valid exclusion restriction, to derive a rule of thumb for the practitioner.
In the remainder of this introduction, the three chapters are briefly outlined.
0.0.2 Evaluating the impact of labor migration on house- hold expenditures using the IV approach
The first chapter uses a standard IV approach to gauge the impact of labor migra- tion on household expenditures in Tajikistan. The effect of labor migration has been the topic of many research projects, and results are often conflicting. Similar to the evaluation of a micro credit program, when estimating the effect of migra- tion one has to take into account the self-selection of individuals into treatment, which is the cause of bias in this case. Rather than being selected into treatment by a program planer, the family’s decision to send a migrant abroad is determined by a number of individual characteristics, both observed and unobserved. The first chapter takes this into account by using an IV approach. Rather than relying on mere intuition, the first stage is based on a simple household-level model determin- ing the migration decision, which is first theoretically derived and then empirically
tested. The needed exclusion restriction is the percentage of migrant households in the cluster of the respective household, which is a standard instrument for estimating the impact of migration. The rationale behind it is straightforward and comes from the idea of migration networks (see, for example, Carrington and Vishwanath, 1996, Bauer et al., 2002, Woodruff and Zenteno, 2007, McKenzie and Rapoport, 2010): The more families in the vicinity have migrants, the easier it is to receive information and send a migrant yourself.
It is often assumed that international labor migration from Tajikistan, while hav- ing no noticeable effects on investment (usually defined as medium and long-term consumption, such as education, or investment into housing or business), on av- erage leads to an increase in short-term consumption, mostly food. While only weak effects of migration measured by a simple dummy are visible, repeating the analysis instead using the length of the migration spell, as well as its squared term reveals that labor migration apparently takes a while to ”kick in” and be prof- itable to those remaining at home. The observed long-term effects on household consumption patterns, albeit being rather small, actually speak in favour of in- vestment of remittances, with the respective shares increasing over time, while the budget share spent on food slowly decreases.
0.0.3 Estimating the effect of a macroeconomic shock using a heterogeneous exposure approach
The second chapter, which is joint work with Antje Kroeger, deals with a some- what different evaluation problem, namely the impact of the recent financial crisis on Tajik labor market dynamics. The challenge here is the obvious lack of con- trol group data, since the crisis hit worldwide. Although this makes an impact evaluation in the true sense impossible, we try to assess the effect of the cri- sis by using a heterogeneous exposure approach. Here we argue that individuals working in the manufacturing sector are comparatively more afflicted by the crisis than those working in other areas, since this part of the economy suffered most. A multinomial probit regression is used to calculate transition probabilities and their determinants between employment categories between 2007 and 2009. A dummy indicating prior employment in the manufacturing sector is included in the model
and interpreted as a proxy for exposure to crisis. Our results suggest a negative impact of the crisis on wage employment, which seems to be somewhat mitigated by labor migration in the rural areas. There, labor migration might indeed be a way of financing labor market exit, while for urban areas this cannot be ob- served. Also, there are no clear indications of the informal sector (measured as self-employment) to act as a shock absorber during the crisis.
0.0.4 A sensitivity analysis of the bivariate probit estima- tor in the presence of distributional misspecifications
The third chapter investigates the performance of the recursive bivariate probit estimator, with and without reliance on a valid exclusion restriction. The aim is to give some advice to the practitioner with regard to the use of instruments. How robust is the functional form assumption (meaning that models are estimated with- out an exclusion restriction) to misspecifications of the error distributions? Can possible bias due to non-normal errors be reduced by the inclusion of valid instru- ments, and, most importantly, could it be exacerbated by the use of faulty (i.e.
endogenous) ones?
By simulating estimation entirely without an instrument (relying on functional form for identification), with a true instrument, as well as with a faulty (endoge- nous) one, some light is shed on this question. Results suggest that departures from normality lead to a noticeable increase in bias, especially if the error distri- bution is highly skewed. Furthermore, the response frequencies of the two binary outcome variables play an important role. Bias tends to increase for unbalanced distributions of the outcomes. With regard to the inclusion of doubtful instru- ments, rather than reliance on functional form, the results suggest that while valid instruments do little to improve estimation in the face of non-normal errors, en- dogenous instruments can noticeably worsen results if the true underlying error distribution is normal. A rough rule of thumb to be derived from this would there- fore be to use the bivariate probit model only when errors can be assumed to be normal, and, in this case, rely on functional form rather than risking increased bias due to faulty instruments. While this advice sounds rather straightforward, the difficult decision as to when errors are normal remains to be made. While tests for
this exist (see, for example, Murphy, 2007), their performance crucially depends on sample size, a feature that can also be observed for the simulations of chapter 3.
Low-skilled labor migration in Tajikistan:
Determinants and effects on expenditure patterns
1.1 Introduction
The impact of remittances from labor migration in developing countries has been the topic of extensive research. The question of how they influence overall poverty outcomes (see, for example, Adams and Page, 2005, Gupta et al., 2009), as well as the income distribution of the recipient country has been of central interest (e.g. Adams, 1989, Barham and Boucher, 1998, Acosta et al., 2008, Shen et al., 2010). However, some studies also take a more micro-oriented approach and in- vestigate the effects on well-being of the households remaining at home. Results of these studies are mixed, some concluding that remittances are mostly used to cover day-to-day needs, rather than being invested productively (see, for exam- ple Lipton, 1980, Orozco et al., 2005, Cohen, 2005, Matthieu, 2011), while others find significant increases in household investments (e.g. Adams, 1989, Adams and Page, 2005, Acosta et al., 2007, Woodruff and Zenteno, 2007). A closely related question in this context are the determinants of labor migration. Since leaving
1
one’s home country is usually quite costly both financially and emotionally, the driving factors have to be substantial to justify such a big step. It is straightfor- ward to assume that unemployment faced at home plays a role, or, if that is not the case, that earning prospects abroad are significantly better. Also, the decision to migrate most likely is not an individual one, but one made at the household level (see, for example Stark, 1984, Taylor, 1987, Kainaiaupuni, 2000). If a house- hold member is to go abroad, the necessary financial means for the journey etc.
have to be available, also there might be a need to reshift responsibilities and the work burden within the household, especially if the family member leaving was previously unemployed and doing chores at home. This might lead to a reduction in labor supply offered domestically, for which the literature finds some evidence.
(see Justino and Shemyakina, 2010 for Tajikistan, as well as Amuedo-Dorantes and Pozo, 2006 and Funkhouser, 1992 for the Latin American context). Finally, the existence of migrant networks seems to play an important role, facilitating orientation in the foreign job market (Carrington and Vishwanath, 1996, Bauer et al., 2002, Woodruff and Zenteno, 2007, McKenzie and Rapoport, 2010).
Somewhat less attention has been given to the question whether there exists a dif- ference between short and long-term effects of remittances from labor migration.
Most of the existing work (see, for example, Taylor, 1992) base the (assumed posi- tive) long-term effects mainly on productive asset accumulation. In this paper, an alternative hypothesis is introduced. It is assumed that the early effects of labor migration might be almost non-existent or in some cases even negative, since the initial costs of migration, which especially for Tajikistan seem substantial, first have to be redeemed by inflowing remittances. This process is slowed down due to the fact that new migrants need some time to establish themselves in the new working environment and find profitable jobs. It is further delayed by a possible reduction in domestic work income due to reduced labor supply of the remaining family members, who have to fill the gap the migrant left behind. The contribution of this paper is therefore three-fold: First, the determinants of low-skilled1 labor
1 Since the majority of labor migrants takes on low-skilled jobs, such as construction work, abroad (even if the amount of people with secondary education is quite high among migrants), the analysis is limited to the effects of such types of employment. It needs to be noted, however, that the inclusion of high-skilled labor migrants does not really change the results obtained, since the number of observations is very low here.
migration in the Tajik context are theoretically modelled and empirically tested.
Using the results to control for the selectivity of migration, the impact of migra- tion on household expenditure shares is then estimated. Finally, this analysis is extended to include the length of the migration spell, rather than a simple dummy or the amount currently remitted, to gauge possible long-term effects. Addition- ally, effects of migration on labor supply are also (albeit tentatively) investigated.
To the author’s best knowledge, this is the first time this is attempted for the case of Tajikistan. However, since the data used are cross-sectional, naturally the analysis of long-term effects is somewhat limited. Further research on this topic is needed, making use of panel data sets with detailed migration information.2
1.2 The case of Tajikistan
The Republic of Tajikistan is a small, landlocked country in Central Asia, the poorest among the states of the former Soviet Union.3 A number of factors make the economic development of Tajikistan problematic. First of all, over 90% of its territory is mountainous, with about 50% as high as (or higher than) 3000 meters above sea level, and only approximately 7% of it suitable for farming.
Natural resources are limited. Both agriculture and industry are almost exclusively centered on cotton production and aluminum,4 a remainder from central planning under the Soviets, leaving the economy very vulnerable to fluctuations in demand for these commodities. To add to this, the country suffered a devastating civil war (1992-1997), following the break-up of the Soviet Union, which also strongly impeded Tajikistan’s economic development. In such a setting, labor migration seems like a natural mitigation strategy, and indeed, Tajikistan has one of the highest (if not the highest) percentage of remittances to GDP.5 The most popular country of destination is Russia, since many Tajiks still have at least a working
2Although the TLSS 2007 used here can be used as a panel with the TLSS 2009, results might be distorted due to the external shock caused by the financial crisis during that period.
3In 2010 it had a HDI of 0.58 and therefore ranked 112th among 169 countries (see UNDP, 2010a).
4Another branch of industry now gaining in importance, but still comparatively small is electricity generation through hydropower.
5Some estimates yield figures as high as 45% in 2008 (see Ratha et al., 2008. A more recent figure is 35%, Yang, 2011).
knowledge of Russian, and are allowed to enter the country without visa.6
1.2.1 Tajik labor migration
While the effect of labor migration on the households at home has been the topic of various research, the case of Tajikistan, which is one (if not the) leading coun- try in terms of labor migration, so far has been somewhat neglected. The (to the author’s knowledge) only paper attempting a methodologically rigorous im- pact evaluation of labor migration on consumption and investment patterns in Tajikistan is the work by Matthieu, 2011. Using propensity score matching on the 2003 Tajikistan Living Standard Survey (TLSS) he finds a positive significant effect of external remittances on per capita food consumption, while ”investment”
expenditures (in his definition those include expenditures on health, education, agriculture, rent, utilities, as well as transfers to others) are negatively affected.
Olimova and Olimov, 2007 reach the same conclusion doing a descriptive analysis of migrant families, with focus on the high-altitude regions of Tajikistan. They assume that remittances from labor migrants are mostly used to cover day-to- day needs, and do not lead to significant capital accumulation or investment. A number of other articles (e.g Mughal, 2007, Olimova and Bosc, 2003) support this theory, albeit without empirically testing it.
1.3 Data
The data source used is the 2007 Tajikistan Living Standard Survey (TLSS 2007), prepared by the World Bank in collaboration with UNICEF and carried out by the National Committee for Statistics (former Goskomstat, now Tajstat). The survey is representative on the national, rural/urban, as well as the district (oblast7) level, with the sampling frame based on the 2000 Census of Tajikistan.
6While entering Russia is easy for Tajik citizens, obtaining legal residence and work permits is often significantly more difficult, giving rise to large amounts of illegal workers and the asso- ciated problems. Especially following the recent financial crisis, Russian immigration laws have tightened, making legal labor migration harder for Tajiks.
7Tajikistan is divided into 5 administrative regions or oblasts: Dushanbe, RRS, Soghd, Khat- lon, and GBAO.
The survey has a complex survey design, with a total of 270 clusters, where each cluster is either fully urban or fully rural and contains 18 households. The total number of households is 4860. Due to missings in some needed variables, the final estimation sample comprises of 4715 households. The survey includes data on the socio-demographic composition of the household, labor market activities, the health and education of individuals, transfers to the household from various sources and a very detailed module on migration.
1.4 Modeling determinants of labor migration
Since labor migration is a highly selective process, further analysis of its impacts needs to be preceded by a solid investigation of its determinants. In the following, a simple, one-period income optimization model on the household level8 is derived, similar to the one used by McKenzie and Rapoport, 2007.
Assume that the household’s disposable income π is given by π=bL∗log(F −K−m)−[(F −m)I−A] +m(w−c) where
• L = Farmable land available to HH
• F = Number of HH members
• K = Number of dependent HH members (either too young/old to work or disabled)
• m = Number of migrants currently abroad
• I = Subsistence cost per HH member
• A = Amount of additional financial means (e.g. transfers from other family members, friends, etc.)
8Sending a migrant away is not only costly, but also has significant implications for family life, such as intra-household organization and sharing of work burden. Therefore it is argued that the decision to migrate is made on the household, rather than on the individual level.
• w = Wage earned by migrant abroad
• c = Cost of sending a migrant abroad
• b = Some parameter (0< b >1)
Now maximizeπwith respect to m, the number of migrants to be sent abroad, subject to the constraint that additional financial meansAmust cover all migration costs c.9
Maxm π, s.t. A≥mc If the restriction is not binding (i.e. λ= 0):
m∗ = −bL
(I+w−c)+ (F −K)
The implications of this equation are straightforward and not surprising: A negative impact of farmable land on labor migration emerges, which makes sense in a predominantly rural country such as Tajikistan. If the household has enough land, more working age family members are needed to farm it. An increase in money needed to send a migrant away has the same effect. The decision to send a migrant is positively influenced by the amount of working age household members, the wage earned abroad, as well as the subsistence costs at home.
If, on the other hand, the restriction is binding, the above equation becomes m∗ =A/c
meaning that the amount of additional means needed to finance the departure of a migrant becomes the bottleneck and therefore the sole determinant of the decision to migrate. In the following, these theoretical results will be empirically
9It can be argued that this is somewhat artificial, since migration could also be financed using, for example, regular income. However, the implicit assumption made here is that migrant households are generally too poor to fully fund migration through income and therefore have to rely on external financing. If it were possible for them to cover the substantial sum needed to send a migrant fully with regular work income, migration most likely would not be economically necessary for the household.
tested, where the estimation serves as the first stage equation of the 2SLS ap- proach to determine the impact of labor migration on expenditure patterns. The following section outlines the econometric strategy in more detail.
1.5 Econometric methodology
When analyzing the effects of labor migration on household-level expenditures, it has to be taken into account that there most likely also exists a reversed causality:
Not only are expenditures influenced by migration, but they might also have an effect on the decision to migrate. A household has to have a certain amount of income in order to be able to afford sending a migrant. Also, if the income situation of the household is already satisfactory, migration might not be needed at all. To account for the endogeneity of labor migration, a 2SLS model is used. Following the related literature on network effects, the percentage of neighboring households with at least one migrant is used as an instrument in both cases.10 As already mentioned in the introduction, the exogeneity of the instrument cannot be tested, and is motivated using the network hypothesis. And indeed it makes intuitive sense to argue that the density of migrants surrounding a household does have an effect on its income situation only through the enhanced chance of sending its own migrant abroad, making use of knowledge and contacts already established by others. The literature also finds evidence in favour of this (see, for example Carrington and Vishwanath, 1996, Bauer et al., 2002, Woodruff and Zenteno, 2007, McKenzie and Rapoport, 2010). Some critics of such cluster-percentage instruments however claim that these variables only reflect regional disparities.
To check for this, the analysis was repeated including oblast dummies. Since they were never significant, one can conclude that systematic regional differences are not problematic here. The lack of significance is also the reason why they are omitted from the results shown in this paper.
To investigate the impact of the length of the migration spell, the 2SLS approach is slightly modified. It is assumed that the endogenous regressor (i.e. the length
10For each household, this value is calculated as the percentage of households with at least one migrant in the respective sampling cluster, excluding the household in question.
of the migration spell in months) enters the estimation equation once linearly, and once as its square, to capture possible reversing effects over time.11
1.6 Descriptives
Before the results of the analysis are presented, a short descriptive overview of the data is given. All figures are estimated proportions within the population. N is the sample size in households or individuals, respectively.
Table 1.1: Proportion of HHs with and without migrants Proportion
No migrant 0.886
At least 1 migrant 0.114
N 4715
Table 1.1 shows the distribution of migration households. In Tajikistan, around 11% of households have at least one migrant. Note that for the purpose of this analysis, only migrants currently abroad, which remit in either cash and/or kind are counted.
Tables 1.2 -1.3 give some more information about personal characteristics of the migrant. As can be seen, most migrants are male (96%), have at least secondary education (84%) and are relatively young, with a mean age of 28.
Table 1.2: Gender distribution among migrants Proportion
Female 0.0421
Male 0.958
N 734
11The first and second stage for this model are estimated manually, and are then bootstrapped with 200 repetitions to receive corrected standard errors.
Table 1.3: Proportion of secondary education or higher among migrants Proportion
No sec. educ. 0.157
Sec. educ. or higher 0.843
N 734
It is also interesting to see that a substantial part (around 65%) of those cur- rently working abroad was unemployed prior to migration (table 1.4), which lends some evidence to the theory that labor migration might be a mitigation strategy for unemployment at home.
Table 1.4: Activity prior to migration Proportion
Working 0.286
Unemployed 0.646
Studying or other 0.0679
N 734
Also, about 80% of migrants come from rural areas, which is not surprising considering the fact that Tajikistan is a predominantly rural country, with only about 32% of the population being classified as urban. As already mentioned, the main country of destination is Russia (97%, see table 1.5):
Table 1.5: Country of destination Proportion
Russia 0.972
Other CIS 0.0108
Rest of the world 0.0167
N 734
1.7 Results
1.7.1 Testing the model
First, the validity of the migration model is tested, which will then serve as the first stage of the 2SLS impact regression. The outcome is a binary variable, which takes the value 1 if the household currently has at least one remitting migrant abroad, and zero otherwise.12 A probit model with the following covariates is fitted:
• farmable land per capita
• the intra-household dependency ratio13
• a dummy indicating access to additional cash (e.g. possibility to borrow from friends/relatives, etc)
• the intra-cluster percentage of households with at least one migrant
• a dummy indicating whether the household head has secondary (or higher) education
• the age of the household head, as well as the age squared
• a dummy indicating whether the household head is currently unemployed
• a continuous variable measuring the altitude
• a dummy indicating whether the household is rural or urban
The first three covariates are directly derived from our model. The intra-cluster percentage of migrant-households is added as a proxy for migration networks. Since it will be used as an instrument in the following analysis, this regression can be seen as a relevance test. Finally, some additional household characteristics are added, in accordance to the related literature. Altitude is assumed to influence the mi- gration decision, since job opportunities are hard to come by in the high-altitude
12In addition to this specification, the model was also tested using a categorical outcome (0, 1, 2 or more migrants). Results are very similar and are omitted here.
13Calculated as hh members younger than 14 or older than 65 working age hh members .
regions of the country, thus increasing the incentive to look for work elsewhere.
In the Tajik context, the altitude variable can also be interpreted as an indicator for general infrastructure, such as transport, proximity to banks or post offices, which become scarce with increasing altitude. Also, a location dummy is included to indicate whether the household is rural or urban.
Looking at the results in table 2.9, we see a confirmation of our model: The lower the relative number of working age household members, the lower the proba- bility to send a migrant abroad (which is strictly logical). Also, a significant posi- tive impact on migration is observed for access to additional financing.14 Farmable land per capita has the expected negative sign, however, the effect seems quite small and is just short of being significant on conventional levels. The network proxy has a strong, positive and significant effect on the probability to send a mi- grant abroad, and therefore meets the relevance criterion of a suitable instrument.
If the household head has at least secondary education, the propensity to migrate is reduced, assuming that the family is relatively wealthy and might not need to send a member abroad to work and remit. The financial pressure of having an unemployed household head increases the chance of having a migrant, which is not surprising, while the age of the household head does not seem to influence mi- gration.15 As already expected, coming from urban areas reduces the probability of having a migrant.
1.7.2 Impact on expenditure shares
The descriptive comparison of mean expenditure shares in table 1.7 shows almost no difference between migrant and non-migrant households. However, the endo- geneity of migration has not yet been controlled for.16
14Note that no endogeneity problem should arise with this variable, since additional financing is defined as coming from outside the household and should therefore not be influenced by the household’s labor migrants.
15The model included also the squared age of the household head. However, since table 2.9 displays the marginal effects (at the mean of continuous variables), this is already accounted for.
16The displayed shares are exclusive categories and add up to 100%. Non-food expenditures comprise clothing, toiletries and other small items for daily use, while utilties are the costs for rent, heating, water and the like. It should be noted that this category does not contain the estimated rent of owned housing.
Table 1.6: Marginal effects of the probit model (1)
land per capita -0.00227 (-1.63)
tajik (d) 0.00600
(0.53)
dep. ratio -0.0390***
(-4.05) access to cash (d) 0.0849***
(4.55) head sec. (d) -0.0366***
(-3.02) migrant hh cluster perc. 0.426***
(16.25)
altitude 0.00000830
(1.36)
head age 0.000476
(0.97)
location (d) -0.0375***
(-3.43) head unemp (d) 0.0449***
(3.20)
N 4715
t statistics in parentheses
* p <0.10, **p <0.05, ***p <0.01
Table 1.7: Average expenditure shares for HHs with and without migrants
(1) (2)
Mean share for non-migrant hhs Mean share for migrant hhs N
food 0.691 0.697 4715
non-food 0.188 0.182 4715
medical exp 0.03 0.039 4715
utilities 0.051 0.051 4715
education 0.049 0.043 3589
As already outlined in the methodology section, this is done with a 2SLS ap- proach, using the intra-cluster percentage of migrant households as instrument.
The estimation equation is specified based on the approach used by Working, 1943 to estimate Engel curves. As in any Engel curve estimation, expenditure shares are linked to total expenditures. A myriad of suggested functional forms for this rela- tionship exist in the literature (See, for example, Prais and Houthakker, 1971 for experiments with various forms). The specification postulated by Working assumes a linear relationship between expenditure shares and the log of total consumption.
As is shown in Deaton and Muellbauer, 1980, such a relationship satisfies the re- quirements of a utility function. The estimation equation is further extended to accomodate possible economies of scale for different household sizes (see Deaton, 1997, p.231), and thus takes the form:
wi =α+β1log(xi/ni) +β2log(ni) +β3mi+γ0zi+
wherewi is the respective expenditure share for household i,xi are total expen- ditures,ni is household size (excluding migrants currently absent),mi is a dummy variable to indicate whether the household has migrants, and zi is a vector of ad- ditional covariates. Following Taylor and Mora, 2006, an alternative specification was also tested, which included an additional interaction term of the migration variable with the log of total expenditures, to allow for migration to also affect the influence of total expenditures on the shares. However, this interaction term
was never significant, which is why the above specification is used instead. This suggests that having a migrant in the household only influences expenditure di- rectly, and that it does not affect the impact of overall income. This is a somewhat surprising result, which contradicts the findings of Adams, 2005 and Taylor and Mora. It is often argued that the effect of labor migration on consumption goes beyond simple income increase, and that, for example, exposure to different goods and lifestyles through the family migrant causes a change in spending decisions of those remaining at home. However, we do not find evidence for this in the Tajik data. A possible explanation for this could be that Tajik labor migration is of- ten seasonal, meaning that migrants frequently return home for longer stays, thus keeping strong ties with their families and hindering immersion into the culture of the host country.
Following the literature, shares are analyzed separately for food, non-food, med- ical, utilities and educational expenditures. This allows an (admittedly some- what crude) distinction between short-term consumption (food and most non-food items) and more long-term spending, which might be regarded as investments. Ed- ucation is the best example here. Medical expenditures could also be viewed as an investment into human health and therefore productivity. Utilities such as fuel for cooking and heating, as well as water and electricity, are probably best categorized as medium-term expenditures. Unfortunately, the consumption aggregate of the TLSS 2007, which is at the basis of this analysis, does not include expenditures on housing such as rent or home improvement, which should be counted as in- vestments, and play an important role in the Central Asian context. The same is true for agricultural expenditures and the purchase of durable assets. Other than spending on housing, these last two categories can be constructed from the data.
However, separate analysis of these shares yielded no significant effects. They are omitted here, since distributions are quite lumpy around zero, with the majority of households claiming no expenditures, which makes the results somewhat doubtful.
The results are shown in table 1.8 below.
As already anticipated by the descriptive results, the effect of the migration dummy17 on expenditure shares does not seem very prominent. The expected ef-
17As a robustness check, all regressions were repeated using remittances per capita, rather than the migration dummy as the treatment variable, as well as a broader migrant definition,
Table 1.8: Results of 2SLS regression on household expenditure shares
(1) (2) (3) (4) (5)
share food share nonfood share med share housing share educ
hh has migrant (d) 0.145 -0.0944** 0.0451 -0.0503 -0.0474*
(1.45) (-2.38) (1.58) (-1.30) (-1.87)
land per capita 0.000183 0.000152 -0.000411*** -0.0000428 0.0000860
(0.43) (0.43) (-3.81) (-0.25) (0.45)
tajik (d) 0.00577 -0.00287 -0.00531 -0.000464 0.000745
(0.56) (-0.40) (-1.24) (-0.11) (0.15)
dep. ratio 0.0159*** -0.00985** 0.00269 -0.00419** -0.00714***
(2.78) (-2.36) (1.50) (-2.31) (-3.03)
access to cash (d) -0.0109 0.00935 -0.00205 0.00486 -0.00206
(-0.78) (0.82) (-0.23) (1.09) (-0.29)
altitude -0.0000124 0.0000200*** -0.00000954*** 0.0000163*** -0.00000134
(-1.39) (3.59) (-3.37) (3.02) (-0.39)
location (d) -0.0167 0.0161** 0.00483 0.00159 0.0128***
(-1.56) (2.37) (1.13) (0.34) (2.65)
head sec. (d) -0.0196** 0.0132** -0.00125 0.00196 0.0107***
(-2.44) (2.09) (-0.36) (0.51) (3.32)
head age squ. 0.00000251 0.0000162 -0.0000109* 0.00000617 0.00000114
(0.15) (1.19) (-1.72) (0.86) (0.16)
head age 0.000182 -0.00199 0.00124* -0.000737 -0.000134
(0.10) (-1.36) (1.81) (-0.92) (-0.17)
head unemp. (d) -0.0176* 0.000425 0.00805** 0.00352 0.00677*
(-1.90) (0.07) (2.27) (0.87) (1.91)
log total pc exp. -0.131*** 0.0707*** 0.0166*** 0.0118*** 0.0341***
(-13.28) (9.21) (3.77) (3.71) (3.63)
log hhsize -0.0683*** 0.0381*** 0.00739** -0.00734** 0.00872**
(-6.95) (5.88) (2.00) (-2.00) (2.08)
cons 1.454*** -0.190*** -0.0947*** 0.0172 -0.136**
(21.35) (-3.60) (-3.39) (0.74) (-2.59)
First stage F-test 82.87 82.87 82.87 82.87 85.22
N 4715 4715 4715 4715 3589
t statistics in parentheses
* p <0.10, **p <0.05, ***p <0.01
fects associated with an increase in wealth, namely a decrease in the expenditure share on food, as well as an increase in the other, less basic categories, cannot be observed. Quite on the contrary, there seem to be significant decreases for both non-food items and, most worryingly, education.18
Before we move on to further investigate the somewhat counterintuitive observed effects of migration on the different expenditure categories, we will have a brief look at the remaining covariates of the analysis. As would be expected, a high depen- dency ratio, meaning that few work-age household members have to support rela- tively many non-work-age individuals significantly increases the food expenditure share, while it reduces all others. The impact of altitude also holds few surprises, however, it is a little more complex. The very small, yet highly significant increase in the non-food share is most likely an artefact of insufficient deflation. As already mentioned, infrastructure strongly deteriorates with increasing altitude, meaning that goods are more expensive due to excessive transportation costs. While this effect is accounted for for food expenditures by using regional price deflators, we cannot fully control for them in the case of non-food items, since the deflators are based on food prices. The same could be true for the increased spendings on utility, however, here the harsher climate with noticeably colder winters could also add to costs. A decrease in medical expenditures can also be explained with lack of infrastructure in the highlands, which makes receiving medical help difficult and probably often leads to self-medication, rather than visiting a facility. All in all, altitude seems to matter, although the magnitude of the effect is quite small.
Whether a household is rural or urban has the expected effects on expenditure shares, however, they are only significant for non-food spending and education. If the household head has secondary or higher education, food expenditure shares are relatively lower, while an increase is observed for non-food, as well as education spending, which is intuitive.
The most puzzling effect is the one observed for an unemployed household head, which seems to lead to a decrease in food expenditure shares, as well as an increase
also including already returned, as well as non-remitting migrants. Results are always equal in sign and usually also in significance.
18To avoid excessive censoring around zero, expenditure shares for education were analyzed on the subsample of households with at least one school age child. ”School age” here is defined as being between 7 and 17 years of age, which in 2007 was the mandatory school age in Tajikistan.
in spending on medication and education. One possible, yet somewhat unlikely, explanation could be that scarce funds are redirected into education, to avoid more household member unemployment in the long run. It could also be that causality is reversed, meaning that poor health leads to unemployment. Since this would bias the coefficient of the dummy indicating unemployment of the household head rather than the coefficient of interest (i.e. that of the migration dummy), it seems safe to ignore this possible endogeneity problem here.
The logarithm of total per capita expenditure is always highly significant and shows the expected sign. With increasing income, households tend to spend less of their resources on food, and more on the other categories. Some economies of scale can be observed for household size, namely for food and utilities, which is what one would expect.19
The question remains as to why so few effects of labor migration on expen- diture shares can be found for low-skilled labor in Tajikistan, and why for some sub-categories there actually seems to be a negative impact. When looking at the results of Matthieu, 2011, who does an analysis of per capita expenditure lev- els for 2003 Tajik data using propensity score matching, one finds this effect at least partly repeated. Although he observes a significant and positive effect on per capita expenditures on short-term consumption (defined as food and non-food items), a significant negative effect of almost equal magnitude emerges for the remaining, more long-term consumption categories emerges, thus leading to an overall effect on expenditures somewhere close to zero.
This paper offers and tests two hypotheses as to why the impact of labor mi- gration on expenditures seems so low. Firstly, as already outlined above, labor migration in Tajikistan possibly needs some time to become profitable for those remaining at home. Secondly, labor migration might cause a reduction in labor supply of those staying behind. In the following, these two possibilities are inves- tigated.
19The huge and highly significant effect of the constant in the regression for food share is somewhat surprising. However, similiar results are obtained for the food share by Adams and Page, 2005 for Guatemala.
1.7.3 Long-term effects of migration
It is easy to imagine that sending a family member abroad to find work warrants some initial costs, which can be substantial, relative to family income. Also, establishing oneself as a worker in a foreign country can take some time, during which returns will be modest, and possibly even negative. Anecdotal evidence for this can be found in Kumo et al., 2011 and Ganguli, 2009, where interviews with Tajik migrant workers in Russia were conducted. Not only are costs of travel rather high, but legal issues, such as work permits also initially take up a lot of resources.
Finding and keeping lucrative work may further be hindered by exploitation by employers, as well as harrassment of migrant workers through Russian officials, which seems to be quite common. So it is easy to imagine that the newly arrived migrant needs some time to install himself in a profitable working place. Also, since a substantial part of migration seems to be seasonal, travel frequency back and forth increases, which naturally also drives costs. A first, descriptive confirmation of this hypothesis is the t-test of the mean monthly amount remitted both by recent and more long-term migrants. ”Recent” here is defined as having been away no longer than 5 months.20 Table 1.9 shows significantly lower mean remittances for new migrants, thus lending first support to the time hypothesis.
Table 1.9: Mean comparison of remittances (in Tajik Somoni) between recent and long-term migrants
recent migrant (<= 5 months) long-term migrant (>5 months) mean difference t-value
85.56 296.2 -210.64*** -7.63
N 734
The t-test was repeated several times, gradually increasing the time span con- sidered ”recent”. Differences between the two groups of migrants seem to disappear around a migration spell of 15 months. To further investigate this, the 2SLS anal- ysis was repeated, using migration spell and migration spell squared (measured in months since departure) as endogenous variables, to see if initial negative effects might be reversed over time. Looking at the results in table 1.10, we see some
20To accomodate seasonal migration, migration spells include returns to home of up to 3 months).
confirmation of the time hypothesis (Since the estimates for the covariates other than the migration spell and its square are very similiar to those in table 1.8, a dis- cussion of them is omitted here). The expenditure share for food first significantly increases, and then starts to slowly decrease with migration duration. Shares for utilities, on the other hand, show long-term growth after initial decrease. There also seems to exist a positive effect of migration time on medical expenditures, however, significance is quite weak here and can only be observed for the interac- tion term. The negative effect of migration on education expenditures apparently is not reversed over time, but actually exacerbated, which is cause for concern. A possible explanation could be negative signalling. Since most Tajik migrants work in low-skilled jobs abroad, although the majority of them has secondary eduction, this might send out the wrong message regarding the future usefulness of school- ing. If even with higher education, working abroad on a construction site is the most lucrative option (which most likely could also be achieved without secondary education), spending money on more than basic schooling seems somewhat point- less. It also needs to be noted, however, that this effect is not robust to alternative estimation samples. If, for example, the definition of ”school age” is broadened, this effect also reverses. From the present results it is therefore not possible to draw a final conclusion concerning the effect of migration on education spending.
For all expenditure categories, the observed effects of the length of the migration spell are quite small, but it has to be kept in mind that duration is measured in months (ranging from 0 for non-migrant households to a maximum of 104). Fi- nally, it goes without saying that longitudinal data are of course needed to fully gauge the intertemporal effects of labor migration on expenditures. Nevertheless, this analysis is a first step in the direction of analyzing such effects and gives some indication that the hypothesis of positive, but delayed effects of migration is valid.
1.7.4 Effects of migration on domestic labor supply
In addition to the time needed for labor migration to become profitable, reduced labor supply of those remaining at home, and therefore reduced domestic income, could also be a reason why we do not immediately observe the expected effects of migration on expenditures. One could think of two main reasons for remaining
Table 1.10: Time effect of migration on household expenditures
(1) (2) (3) (4) (5)
share food share nonfood share med share housing share educ land per capita 0.000368 0.000190 -0.000491*** -0.000151 0.0000252
(0.90) (0.56) (-4.93) (-0.88) (0.14)
tajik (d) 0.00298 -0.00235 -0.00505 0.000864 0.00150
(0.43) (-0.41) (-1.43) (0.28) (0.39)
dep. ratio 0.0180*** -0.00920** 0.00159 -0.00544** -0.00746***
(3.26) (-2.23) (0.71) (-2.45) (-3.13)
access to cash (d) -0.0116 0.00874 -0.00135 0.00537 -0.00228
(-0.96) (0.79) (-0.19) (1.06) (-0.32)
altitude -0.0000172*** 0.0000236*** -0.0000114*** 0.0000179*** 0.000000211
(-3.02) (5.05) (-4.63) (5.04) (0.08)
location (d) -0.0173** 0.0187*** 0.00274 0.00117 0.0134***
(-2.57) (3.80) (0.98) (0.35) (3.39)
head sec. (d) -0.0160* 0.0128* -0.00181 0.000164 0.00973**
(-1.86) (1.94) (-0.50) (0.04) (2.50)
head unemp (d) -0.0184*** -0.00129 0.00973*** 0.00438 0.00694*
(-2.64) (-0.25) (2.78) (1.26) (1.75)
head age sq. 0.00000418 0.0000184 -0.0000132** 0.00000468 0.00000210
(0.28) (1.53) (-2.25) (0.65) (0.25)
head age -0.000107 -0.00220* 0.00150** -0.000526 -0.000209
(-0.07) (-1.71) (2.32) (-0.70) (-0.22)
log pc total exp. -0.134*** 0.0716*** 0.0166*** 0.0131*** 0.0354***
(-18.96) (10.79) (4.37) (4.42) (4.34)
log hh size -0.0680*** 0.0390*** 0.00649* -0.00777** 0.00905**
(-8.45) (6.05) (1.84) (-2.51) (2.09)
migration spell 0.0129*** -0.00237 -0.00124 -0.00615*** -0.00321**
(4.38) (-1.01) (-0.82) (-3.31) (-2.16)
migration spell sq -0.000775** -0.000151 0.000330* 0.000451* 0.000144
(-2.38) (-0.60) (1.64) (1.89) (1.13)
cons 1.471*** -0.197*** -0.0929*** 0.0101 -0.141***
(26.53) (-4.28) (-3.52) (0.41) (-2.80)
First-stage F test 49.8 49.8 49.8 49.8 41.02
N 4715 4715 4715 4715 3589
t statistics in parentheses
* p <0.10, **p <0.05, ***p <0.01
household members to supply less labor domestically. The first is that the house- hold misjudges the new income situation by simply overestimating the expected returns from migration. Less labor is supplied, since it is assumed that future remittances will overcompensate the foregone domestic income. However, such irrational behaviour seems somewhat unlikely. A second, more rational hypothesis is that the reduction in work income is caused by a reshuffling of labor inside the household. Especially if the migrant was unemployed prior to departure (which seems to be the majority of cases, as can be seen in table1.4), most likely he or she was doing some unpaid task at home. A replacement now has to be found among family members, which might lead to a reduction in working hours offered. There is some support for this in the literature (see, among others, Amuedo-Dorantes and Pozo, 2006 and Funkhouser, 1992 for the Latin American context). A recent paper by Justino and Shemyakina, 2010, also confirms this finding for Tajikistan, observing a reduction in work hours for members of migrant households. The same is true for the findings of chapter 2, although it is unclear if the observed effect might be exacerbated by the financial crisis.
To see whether migration indeed has an adverse effect on labor market participa- tion of household members at home, the above analysis from table 1.8 is repeated, using the log of last month’s per capita work income21 as dependent variable. Ad- mittedly, the per capita work income can only serve as a rather crude proxy for labor supply, however, it is the best the data permit. Nevertheless, when looking at the results in table 1.11, the analysis confirms the findings of Justino and She- myakina, who use more detailed data containing information on the hours worked of each household member.
A strong and highly significant negative effect of migration on domestic per capita income can be observed, which is consistent throughout all different speci- fications of migration (as already mentioned, only results using the above defined migration dummy are shown here). Thus the hypothesis of reduced labor supply due to migration is confirmed. (Another interesting result is the fact that ap- parently access to occasional additional funds is not enough to cause significant reduction in labor supply). Arable land per capita significantly reduces work in-
21Per capita figures are calculated excluding migrants currently abroad.
Table 1.11: Results of 2SLS regression on household pc income log dinc pc
hh has migrant (d) -2.734***
(-3.82) land per capita -0.0176***
(-4.28)
tajik (d) 0.00174
(0.02)
dep. ratio -0.375***
(-7.59) access to cash (d) 0.113
(0.68)
altitude -0.000501***
(-5.83)
location (d) 0.0776
(0.74) head sec. (d) 0.133 (1.26) head age sq 0.0000120
(0.08)
head age 0.0240
(1.54) head unemp. (d) -1.481***
(-15.31)
log hh size 0.152*
(1.95)
cons 3.083***
(7.73) First-stage F test 30.7
N 4715
t statistics in parentheses
*p <0.10, **p <0.05, *** p <0.01
come, which is straightforward, since the bigger the family plot, the more people are needed to farm it. As expected, a high dependency ratio, meaning that the household has relatively few work-age members, also decreases family work income.
The same is true for an unemployed household head, which is strictly logical. (As a robustness check, this covariate was omitted from the analysis, however, this did not lead to any changes with regard to the effect of migration on household work income). The small, yet highly significant negative effect of altitude is also not surprising. As already mentioned many times, infrastructure and employment opportunities grow scarce with increasing altitude. Finally, the positive effect of household size is also to be expected: The more household members, the bigger the probability that some of them are of work-age and earning income.
A third possible cause for the observed effects of migration on expenditure (or the lack thereof) exists, which is also connected to labor supply. It could be that remittances are used to start up small enterprises at home. This has been observed for other countries (see, for example, Amuedo-Dorantes and Pozo, 2006 for Mexico, as well as Funkhouser, 1992 for Nicaragua). However, anecdotal evidence (see Mughal, 2007 and Olimova and Olimov, 2007), as well as the results from chapter 2 speak against this hypothesis for Tajikistan.22 Unfortunately the TLSS 2007 data do not provide enough information to fully research this question, so this is left for further research.
1.8 Conclusion
In this chapter, I tried to shed some light on the impact of low-skilled labor migra- tion on household expenditure shares. Results suggest that the impact is rather small. Consumption patterns in Tajikistan apparently are not influenced by migra- tion per se, but by a change in disposable household income. There seem to exist two effects, working in combination, which cause the expected positive effects of migration on expenditure to appear less prominently. Tajik migrants just starting work abroad usually need some time to install themselves in profitable positions,
22Note however, that the results from chapter 2 are observed during the financial crisis. It is therefore unclear whether they can be generalized to hold also for non-crisis times.
which can be seen in the continuous increase in average remittances sent home over time. At the same time, labor supply in the family decreases, most likely due to a reshuffling of responsibilities inside the household. Since the majority of migrant workers were unemployed prior to their departure, their place at home will have to be filled by some other household member, who will then have less time to supply to the labor market. The combination of these two effects leads to the observed initial deterioration in household expenditure patterns, with ris- ing food shares and decreasing shares for non-food items, education and utilities, which are usually associated with lesser wealth. However, with increasing length of the migration spell, these findings at least partly reverse to yield the expected re- sults, namely more money spend on medical services and utilities (which could be counted as medium term or investment-type expenditures), while the expenditure share on food decreases. The long-term effect of migration on education remains unclear. Results actually indicate a worrying decrease over time, however, they are somewhat sensible to the sample, and vary with the chosen definition of ”school age”. Further research using longitudinal data is needed to explore the intertem- poral effects of migration in general, and with respect to education in particular.
In addition to this it would also be interesting to repeat the analysis with more comprehensive expenditure categories, including, for example, money spend on home improvement, which plays an important role in the Central Asian context.
Also, the role remittances play in investment into start ups would be an interesting topic for further investigation.
Employment and the financial crisis
in Tajikistan
2.1 Introduction
Since the economic crisis hit in fall 2008 the world has seen one of the worst economic turmoils in history, including both developed and developing countries.
There is scant information available concerning the impact of the economic crisis on transition economies. This is especially true for the former Soviet Union economies in Central Asia. As one of the poorest regions in the world, Central Asia was dramatically hit by the financial crisis (ICG, 2010, UNDP, 2010b, Lukashova and Makenbaeva, 2009). As remittance-dependent countries, not only did these countries experience decreasing remittance inflows, there were also tremendous changes in domestic labor markets (ILO, 2010, Tiongson et al., 2010). However, the datasets necessary for more in-depths investigation are scarce, since panel data are necessary to examine dynamics of the financial crisis in the labor market.
In the case of Tajikistan, we can draw on unique panel data, using the 2007 and 2009 Tajik Living Standards Measurement Surveys (henceforth TLSS 2007
& 2009) to examine individual labor market and household migrant decisions1 in
1As could already been seen in chapter 1, labor migration plays an important role in the Tajik
25
